Emulating CMAQ using deep learning: A comparative study on simulating surface NO₂, O₃, and PM_2.5 over the CONUS using...

We developed 2D convolutional neural network (2D CNN)-based emulators that efficiently replicate daily mean surface concentrations of NO₂, O₃, and PM_2.5 over the contiguous United States at 12 km resolution for the years 2015-2019. Near-surface meteorology, emissions, and land surface data from the EPA's Air QUAlity TimE Series (EQUATES) dataset were used as input variables, with corresponding CMAQ outputs serving as targets. Leveraging the U-Net architecture, we progressively expanded the range of input features for the NO₂, O₃, and PM_2.5 emulators, to better reflect the increasing complexity of pollutant dynamics and capture species-specific behaviors. CMAQ-simulated and the emulated NO₂, O₃, and PM_2.5 concentrations showed close agreement, with indices of agreement (IOA) reaching up to 0.95, 0.88, and 0.85, respectively. Seasonal performance and city-scale evaluations also confirmed the emulator's ability to reproduce CMAQ simulations under diverse meteorological conditions and evolving emission patterns, with consistent spatiotemporal agreement. The emulator achieved the highest performance for NO₂, with decreasing accuracy for O₃ and PM_2.5. This trend likely reflects the increasing chemical complexity of O₃ and PM_2.5 and their dependence on vertical atmospheric processes, which are not fully represented in the 2D input features, highlighting the need for 3D emulators to better capture vertical dynamics. For a single timestep NO₂ simulation over CONUS, the emulator ran 1064× faster than CMAQ using a CPU-GPU setup. These results demonstrate our deep learning-based approach's advantage for efficient air quality assessment at a simulation accuracy comparable to traditional chemical transport model outputs, while operating several orders of magnitude faster.